Abstract
Humor is an important factor in communication skills. In natural language processing, various attempts have been made to train models in humor through humor classification, recognition, and detection tasks. However, the difficulty of collecting humorous sentences and the subjectivity of humor is among the many limitations of developing models.
As there is currently no existing Korean humor dataset, this study introduced the first Korean humor dataset, KoHaHa. We collected ‘썰(Ssul)’ genres appearing in online and attempted to solve the problem of subjectivity in humor by applying an approach to humor labeling that differed from other datasets. We also used KoHaHa to explore the possibility of humor transferability to large language models through GPT-3.5 and fine-tuning. The results of the A/B test showed that the fine-tuned model with the KoHaHa dataset learned Korean humor forms. However, GPT-3.5 was evaluated to be high for humorous meaning relevance, confirming the limitations of humor transference. |